Back to all posts
Plant Data7 min read·November 1, 2026

How Do You Detect Sensor Drift in Plant Data?

TB
ThynkBlox Automation Team
Automation Engineering

To detect sensor drift in manufacturing, compare the suspect reading with a suitable reference under matched operating conditions, then review how the difference changes over time. Use redundant measurements and calibration records to investigate the disagreement, rather than assuming the sensor is wrong. Mark affected data as suspect while the investigation is open, and carry that status into dashboards and reports.

What should count as suspected sensor drift?

The International Vocabulary of Metrology defines instrumental drift as a continuous or incremental change over time in an instrument's indication, caused by changes in the instrument itself rather than in the quantity being measured.

For an investigation, look for a growing disagreement with a reference, repeated checks that move in one direction, or readings that no longer match an established relationship under comparable conditions.

A reading staying within its normal operating range is not sufficient evidence of accuracy. Equally, a slow change in the trend is not sufficient evidence of drift: the process itself may have changed.

Start with a question you can test: “Does this measurement disagree with an independent check when both represent the same process condition?”

What should you check before blaming the sensor?

Trace the reading from the instrument to the screen. At each stage, compare the value, unit, timestamp and quality status.

Check:

  • The instrument identity and measurement location.
  • The configured range and engineering unit.
  • Scaling in the controller and downstream systems.
  • Whether the displayed value is live, held or calculated.
  • Changes to configuration, equipment or operating conditions.

Resolve unit and scaling questions before evaluating drift. The guide to standardising engineering units in plant data provides a useful starting point.

Also check whether the observations represent the same moment. For timestamp checks across systems, see aligning timestamps across plant data systems.

Keep the original trend and configuration evidence before making changes. Record each finding so the investigation can distinguish a measurement problem from a data-handling problem.

How should you take a reference measurement?

Choose a reference suitable for the measurement range, installation and accuracy needed for the decision. Ask the person responsible for calibration to confirm its suitability and current calibration status.

Write down the comparison method before collecting readings:

Comparison detailWhat to record
Measurement pointWhere the installed sensor and reference measure
Operating conditionProduct, load, process stage and relevant equipment state
TimingWhen each reading was taken and whether conditions were stable
ReferenceInstrument identity, calibration status and stated uncertainty
Decision ruleAllowed disagreement, who approved it and what happens if exceeded

Do not set the acceptance limit from the difference you happen to observe. Agree it from the process requirement and the measurement capability.

Repeat the comparison at relevant operating conditions. A check at one point should support a conclusion about that point; require further evidence before extending the conclusion across the full working range.

Use approved access and isolation procedures. If the measurement supports a safety function, involve a qualified assessor before changing settings, substituting signals or deciding how operation should continue.

What can redundant readings tell you?

Treat a second sensor as comparison evidence, with its own checks to complete.

Ask whether the two readings measure the same physical condition. Confirm their locations, response settings and timestamps before treating a difference as an error.

Then establish their independence. Two agreeing readings may share a source, configuration error or other common influence; agreement alone does not establish accuracy.

For each redundant reading, record:

  • Whether it comes from a separate physical instrument.
  • Which signal path and configuration it uses.
  • Its maintenance and calibration history.
  • Whether a separate reference check supports it.

When two instruments disagree, avoid choosing whichever reading looks more familiar. Use an appropriate reference or controlled check to resolve the disagreement.

How do calibration records help establish the affected period?

Request the results before adjustment and after adjustment, often recorded as “as found” and “as left”.

For each calibration or verification, look for the instrument identity, test points, reference used, acceptance limits and any adjustment. Include replacements, range changes and repairs in the same timeline.

Use those records to answer separate questions:

  1. What evidence supports the instrument's condition at each check?
  2. When did the first unresolved disagreement appear?
  3. Which records and decisions need review between those points?

If an instrument failed its latest check, do not automatically classify every reading since the previous check as wrong. Mark the interval for review and document what remains uncertain.

Avoid drawing a straight correction line between two calibration results unless an approved technical assessment supports that method. Where the evidence cannot establish when the error began, retain that uncertainty.

What would a worked investigation look like?

Illustrative example: a tank temperature reading remains close to its usual operating value, but maintenance raises a concern after an independent check.

The team compares readings during matched, stable conditions using a reference approved for the check.

CheckInstalled readingReference readingDifference
First comparison80.4 °C80.1 °C+0.3 °C
Later comparison80.8 °C80.1 °C+0.7 °C
Latest comparison81.3 °C80.2 °C+1.1 °C

These illustrative readings justify investigating the increasing disagreement. They do not establish its cause or provide a universal pass limit.

The team checks the installation, configuration and reference evidence against the agreed acceptance rule. It marks the unresolved interval as suspect and identifies reports that used the reading.

After corrective work, the team repeats the approved checks. Historical values retain their investigation status unless a documented review supports a different classification.

How should dashboards handle suspect data?

Give suspect data a visible status and an agreed reporting rule.

A useful record includes the measurement identity, affected interval, reason for concern, supporting evidence, responsible person and next action. Keep the original values available.

Carry the status through each destination:

  • Live displays: show a text label beside the value.
  • Trends: identify the affected interval and explain the marker.
  • Calculations: state whether suspect readings are included, excluded or replaced.
  • Exports: include the quality status and relevant explanation.
  • Reports: disclose affected results and the treatment applied.

Do not rely on colour alone. A reader should understand the status from a label or note.

If a replacement reading is approved, label it as substituted and record its source. Define who can release future data from suspect status and what evidence they must review.

What mistakes should you avoid?

Adjusting before recording the discrepancy. Capture the original condition and comparison results first.

Treating a related process value as a reference. Use the relationship to identify a concern, then obtain suitable measurement evidence.

Comparing unlike conditions. Match location, timing and operating state before interpreting differences.

Using an arbitrary alarm threshold. Have the measurement owner approve limits appropriate to the process and reference capability.

Removing inconvenient readings. Preserve the values and explain their status.

Closing the issue after replacement alone. Verify the replacement, review the historical interval and confirm that affected dashboards and reports apply the agreed treatment.

What should the investigation handover contain?

Before closing the work, confirm that the record contains:

  • The original concern and affected measurement.
  • Reference and redundant-reading evidence.
  • Relevant calibration and configuration records.
  • The cause found, or the uncertainty still unresolved.
  • Corrective work and verification results.
  • The suspect interval and reporting treatment.
  • The person accepting the result and the next review action.

Keep the investigation open if the cause remains unresolved. Assign a next action and define how the reading may be used in the meantime.

Frequently asked questions

Can a normal-looking trend rule out drift?

No. Use a suitable reference comparison and supporting records before accepting accuracy. Treat the trend's appearance as context.

How often should reference checks be performed?

Set the interval with the measurement owner using the consequence of error, operating conditions, instrument guidance and previous results. Review it after a failed check or significant change.

Should suspect readings be deleted?

Preserve them with their quality status. If reporting excludes or replaces them, document that decision and make the treatment visible.

Can a dashboard automatically correct the reading?

Require an approved correction method, a defined range of application and verification evidence. Preserve the raw reading and identify corrected values wherever they are used. Our guide to keeping instrument calibration records that support plant data shows how to link that evidence to the historian.

Next step

Want to apply this to your plant?

Every project starts with a fixed-fee assessment, and you can stop after the report.